Begin
Causal self-attention is the mechanism underpinning most of the advances in AI since 2017. In this article, I will step through the computation and hopefully gain a better intuition of how it works.
$$ \text{SelfAttention}(\mathbf{Q}, \mathbf{K}, \mathbf{V}) = \text{softmax}\left( \text{mask} \left(\frac{\mathbf{Q} \mathbf{K}^T}{\sqrt{d}}\right) \right) \mathbf{V} $$At a high level, this function takes one sequence and transforms it into another. A sequence is a list of token embeddings, a tensor of shape $L \times d$, where $L$ is the input sequence length and $d$ is the embedding dimension. Each row of this matrix corresponds to one input token, which is represented as a $d$-dimensional vector.
So why then, are there 3 inputs to $\text{SelfAttention}$? This is because, in the Transformer architecture, the input sequence is projected by 3 different $d \times d$ linear layers. If $\mathbf{X}$ is the input sequence,
$$ \mathbf{Q} = \mathbf{X}\mathbf{W_Q}, \quad \mathbf{K} = \mathbf{X}\mathbf{W_K}, \quad \mathbf{V} = \mathbf{X}\mathbf{W_V} $$where $\mathbf{W_Q}, \mathbf{W_K}, \mathbf{W_V}$ are $d \times d$. So, $\mathbf{Q},\mathbf{K},\mathbf{V}$ are simply different representations of the same input sequence.
Let’s compute $\text{SelfAttention}$ step-by-step. First, we do $\mathbf{Q}\mathbf{K}^T$, which is a $L \times d$ by $d \times L$ dot product, resulting in an $L \times L$ output. What does this do?
$$ \begin{align*} \mathbf{Q} \mathbf{K}^T = \begin{bmatrix} \mathbf{q}_1 \\ \mathbf{q}_2 \\ \vdots \\ \mathbf{q}_L \end{bmatrix} \begin{bmatrix} \mathbf{k}_1^T & \mathbf{k}_2^T & \cdots & \mathbf{k}_L^T \end{bmatrix} = \begin{bmatrix} \mathbf{q}_1 \mathbf{k}_1^T & \mathbf{q}_1 \mathbf{k}_2^T & \cdots & \mathbf{q}_1 \mathbf{k}_L^T \\ \mathbf{q}_2 \mathbf{k}_1^T & \mathbf{q}_2 \mathbf{k}_2^T & \cdots & \mathbf{q}_2 \mathbf{k}_L^T \\ \vdots & \vdots & \ddots & \vdots \\ \mathbf{q}_L \mathbf{k}_1^T & \mathbf{q}_L \mathbf{k}_2^T & \cdots & \mathbf{q}_L \mathbf{k}_L^T \end{bmatrix} \end{align*} $$The result of $\mathbf{q}_i \mathbf{k}^T_j$ is a scalar ($1 \times d$ dot $d \times 1$), and it is the vector dot-product between $\mathbf{q}_i$ and $\mathbf{k}_j$. If we remember the formula
$$ \mathbf{a} \cdot \mathbf{b} = \|\mathbf{a}\| \|\mathbf{b}\| \cos \theta $$we see that the dot-product is positive when $\theta$, the angle between $\mathbf{a}$ and $\mathbf{b}$, is close to 0º and negative when the angle is 180º, or when they point in opposite directions. We can interpret the dot product as a similarity metric, where positive values indicate similar vectors, and negative values indicate the opposite.
So our final $L \times L$ matrix is filled with similarity scores between every pair of $\mathbf{q}$ and $\mathbf{k}$ tokens. The result is divided by $\sqrt{d}$ to prevent the variance from exploding for large embedding dimensions. See Appendix for details.
The next step is to apply the $\text{mask}$ function, which sets all values that are not in the lower-triangular section of the input matrix to $-\infty$.
$$ \text{mask}\left(\frac{1}{\sqrt{d}} \mathbf{Q}\mathbf{K}^T\right) = \frac{1}{\sqrt{d}} \begin{bmatrix} \mathbf{q}_1 \mathbf{k}_1^T & -\infty & -\infty & \cdots & -\infty \\ \mathbf{q}_2 \mathbf{k}_1^T & \mathbf{q}_2 \mathbf{k}_2^T & -\infty & \cdots & -\infty \\ \mathbf{q}_3 \mathbf{k}_1^T & \mathbf{q}_3 \mathbf{k}_2^T & \mathbf{q}_3 \mathbf{k}_3^T & \cdots & -\infty \\ \vdots & \vdots & \vdots & \ddots & \vdots \\ \mathbf{q}_L \mathbf{k}_1^T & \mathbf{q}_L \mathbf{k}_2^T & \mathbf{q}_L \mathbf{k}_3^T & \cdots & \mathbf{q}_L \mathbf{k}_L^T \end{bmatrix} $$To this, we apply $\text{softmax}$, which converts each row of values in the matrix into a probability distribution. The function is defined as a mapping from $\mathbb{R}^L \to \mathbb{R}^L$, where the $i$th output element is given by
$$ \text{softmax}(\mathbf{x})_i = \frac{e^{x_i}}{\sum_{j=1}^L e^{x_j}} \quad \text{for } i = 1, 2, \ldots, L $$Two things to note here:
- The sum of all output elements is $1$, as is expected for a probability distribution
- If an input element $x_i$ is $-\infty$, then $\text{softmax}(x)_i = 0$
After applying the $\text{softmax}$ function to the masked similarity scores, we obtain:
$$ \mathbf{S} = \text{softmax}\left(\text{mask}\left(\frac{1}{\sqrt{d}} \mathbf{Q} \mathbf{K}^T\right)\right) = \begin{bmatrix} S_{1,1} & 0 & 0 & \cdots & 0 \\ S_{2,1} & S_{2,2} & 0 & \cdots & 0 \\ S_{3,1} & S_{3,2} & S_{3,3} & \cdots & 0 \\ \vdots & \vdots & \vdots & \ddots & \vdots \\ S_{L,1} & S_{L,2} & S_{L,3} & \cdots & S_{L,L} \end{bmatrix} $$Where the entries $S_{i,j}$ are defined as:
$$ S_{i,j} = \frac{e^{\text{mask}\left(\frac{\mathbf{Q} \mathbf{K}^T}{\sqrt{d}}\right)_{i,j}}}{\sum_{k=1}^{L} e^{\text{mask}\left(\frac{\mathbf{Q} \mathbf{K}^T}{\sqrt{d}}\right)_{i,k}}} $$The resulting matrix $\mathbf{S}$ has probability distribution rows of length $L$. The final step is to map our value matrix $\mathbf{V}$ by these probability distributions to give us our new sequence.
$$ \begin{align*} \text{SelfAttention}(\mathbf{Q},\mathbf{K},\mathbf{V}) &= \mathbf{S}\mathbf{V} \\ &= \begin{bmatrix} S_{1,1} & 0 & 0 & \cdots & 0 \\ S_{2,1} & S_{2,2} & 0 & \cdots & 0 \\ S_{3,1} & S_{3,2} & S_{3,3} & \cdots & 0 \\ \vdots & \vdots & \vdots & \ddots & \vdots \\ S_{L,1} & S_{L,2} & S_{L,3} & \cdots & S_{L,L} \end{bmatrix} \begin{bmatrix} \mathbf{V}_1 \\ \mathbf{V}_2 \\ \mathbf{V}_3 \\ \vdots \\ \mathbf{V}_L \end{bmatrix} \\ &= \begin{bmatrix} \begin{array}{l} S_{1,1} \mathbf{V}_1 \\ S_{2,1} \mathbf{V}_1 + S_{2,2} \mathbf{V}_2 \\ S_{3,1} \mathbf{V}_1 + S_{3,2} \mathbf{V}_2 + S_{3,3} \mathbf{V}_3 \\ \hspace{2.4cm} \vdots \\ S_{L,1} \mathbf{V}_1 + S_{L,2} \mathbf{V}_2 + \cdots + S_{L,L} \mathbf{V}_L \\ \end{array} \end{bmatrix} \end{align*} $$Note that $S_{i,j}$ is a scalar, and $\mathbf{V}_k$ is a $1 \times d$ embedding vector. Visually, we observe that SelfAttention is selectively combining Value tokens, weighted by a probability distribution generated by how well the queries and keys attend to each other, i.e. have a large inner product. We also see the weight of an output token at index $i$ is dependent only on the input tokens with index $\le i$, due to the causal mask we applied earlier. This is based on the causal assumption, that the an output token $\mathbf{O}_i$ does not depend on future tokens, which is required when training autoregressive (i.e. next token prediction) models.
Hopefully you found this helpful!
Appendix
Why Scale by $\sqrt{d}$?
We do this to keep the variance from exploding as $d$ increases.
Assume that $\mathbf{q}_i, \mathbf{k}_i \sim \mathcal{N}(\mu = 0, \sigma^2 = 1)$ and i.i.d. Let’s compute the mean and variance of the unscaled $s = \mathbf{q} \cdot \mathbf{k}$.
The mean is trivially zero:
$$ \mathbb{E}[s] = \mathbb{E}\left[ \sum_{i=1}^d \mathbf{q}_i \mathbf{k}_i \right] = \sum_{i=1}^d \mathbb{E}[\mathbf{q}_i \mathbf{k}_i] = \sum_{i=1}^d \mathbb{E}[\mathbf{q}_i] \mathbb{E}[ \mathbf{k}_i] = 0 $$And the variance is:
$$ \text{Var}(s) = \mathbb{E}[s^2] - (\mathbb{E}[s])^2 = \mathbb{E}[s^2] = d $$because
$$ \mathbb{E}[s^2] = \mathbb{E}\left[ \sum_{i=1}^d \sum_{j=1}^d \mathbf{q}_i \mathbf{k}_i \mathbf{q}_j \mathbf{k}_j \right] = \sum_{i=1}^d \sum_{j=1}^d \mathbb{E}[\mathbf{q}_i \mathbf{k}_i \mathbf{q}_j \mathbf{k}_j] $$which is $0$ for $i \ne j$ (since $\mathbf{q}_i,\mathbf{q}_j$ and $\mathbf{k}_i,\mathbf{k}_j$ are i.i.d). For $i=j$,
$$ \sum_{i=1}^d \mathbb{E}[\mathbf{q}_i^2 \mathbf{k}_i^2] = \sum_{i=1}^d \mathbb{E}[\mathbf{q}_i^2] \mathbb{E}[\mathbf{k}_i^2] = \sum_{i=1}^d 1 \cdot 1 = d $$since $\mathbb{E}[\mathbf{q}_i^2] = \mathbb{E}[\mathbf{k}_i^2] = \sigma^2 = 1$.
So if we scale by $1/\sqrt{d}$, our new variance is
$$ \text{Var}\left(\frac{s}{\sqrt{d}}\right) = \frac{1}{d} \text{Var}(s) = 1 $$as desired.
Multi-Head Attention
Most modern systems use multi-head attention, which computes $\text{SelfAttention}$ in parallel over several “heads”. We usually let $d_k=d_v= d_{\text{model}} / H$, where $H$ is the number of heads.
$$ \begin{align} \mathbf{Q}_h &= \mathbf{X} \mathbf{W}^Q_h \quad &\mathbf{W}^Q_h \in \mathbb{R}^{d_{\text{model}} \times d_k} \\ \mathbf{K}_h &= \mathbf{X} \mathbf{W}^K_h \quad &\mathbf{W}^K_h \in \mathbb{R}^{d_{\text{model}} \times d_k} \\ \mathbf{V}_h &= \mathbf{X} \mathbf{W}^V_h \quad &\mathbf{W}^V_h \in \mathbb{R}^{d_{\text{model}} \times d_v} \end{align} $$ $$ \text{head}_h = \text{SelfAttention}(\mathbf{Q}_h, \mathbf{K}_h, \mathbf{V}_h) = \text{softmax}\left( \text{mask} \left( \frac{\mathbf{Q}_h \mathbf{K}_h^T}{\sqrt{d_k}} \right) \right) \mathbf{V}_h $$ $$ \begin{aligned} \text{MultiHead}(\mathbf{Q}, \mathbf{K}, \mathbf{V}) &= \text{Concat}(\text{head}_1, \text{head}_2, \ldots, \text{head}_H) \end{aligned} $$